Darshil Shah
← Back to home13 projects · 3 live

All projects.

Three products you can open and use today, then the machine-learning and LLM systems I’ve built at work, and my research.

01Live

Patient Recap

LLM Pipeline · Healthcare · Full-stack

The last consultation, in ninety seconds.

A web app for doctors: record or upload a consultation and get back a structured clinical summary, a care plan and a follow-up reminder, ready before the patient's next visit.

  • Record in the browser or upload audio; speech-to-text with speaker labels turns it into a doctor / patient transcript
  • A three-step LLM chain extracts what was said, analyses it, then writes a brief the doctor can read in 90 seconds
  • “Come back in three weeks” becomes a dated follow-up in a queue, which closes itself when the next visit is booked
  • Python
  • FastAPI
  • LangChain
  • Gemini / Bedrock
  • Deepgram
  • Postgres (Neon)
  • Next.js
Try it liveGoogle sign-in
02Live

Talent Insight

Recruiting · Voice AI · Evidence-based scoring

Every match comes with its evidence.

A recruiting assistant: pick a job, upload a resume and see how well the candidate fits, with the exact resume lines behind every judgment. Then an AI interviewer can screen the candidate by voice or text and write a report for the recruiter.

  • Every job requirement is marked met, transferable, partial, not met or unknown, each with a quote from the resume
  • A real-time AI voice interviewer (Deepgram voice agent with Claude on Amazon Bedrock), plus a text-chat alternative
  • The screening report backs every judgment with the candidate's own words, and the hiring decision stays with a human
  • Python
  • FastAPI
  • LangChain
  • Amazon Bedrock
  • Deepgram Voice Agent
  • Next.js
  • Postgres (Neon)
Try it liveNo sign-in needed
03Live

AI Catalog Intelligence

Document AI · Vision · Async pipeline

Messy supplier files in, a clean catalog out.

Upload a supplier's messy files (Excel sheets, PDF spec sheets, product photos) and get a clean, structured, validated product catalog.

  • Reads spreadsheets, PDFs, scanned pages (OCR), images and ZIP folders; an LLM maps each supplier's columns
  • Matches product photos to the right SKU using codes, filenames, OCR text and image embeddings; uncertain matches go to a human review screen
  • Writes titles, descriptions and bullet points from verified data only, with every value traceable to its source file and row
  • Python
  • FastAPI
  • Amazon Bedrock
  • Titan Embeddings
  • pgvector
  • RapidOCR
  • OpenCV
  • AWS EC2 / S3
  • Next.js
Try it liveGoogle sign-in · limited access
04Aviso AI

Context-Aware Summarization

GenAI · Conversation intelligence

Every call summarised for what it actually was.

An 11-node LangGraph agentic pipeline that classifies calls across Sales, Post-Sale, Internal and Outbound, then routes each to the LLM workflow that fits it.

  • LLM-driven participant role assignment and call-type / sub-type classification
  • Brings CRM, participant and revenue-intelligence context into the summary
  • Deployed through TrueFoundry and AWS ECR with LLM tracing for observability
  • LangGraph
  • LangChain
  • LLMs
  • AWS
  • TrueFoundry
05Aviso AI

Smart Trackers

LLM · Human-in-the-loop

Custom conversation trackers from a handful of phrases.

A few-shot LLM framework that lets clients define their own conversation trackers by annotating semantically similar sentences from past calls.

  • Retrieves similar historical sentences from AstraDB for the client to mark Yes / No
  • Builds tracker-specific positive and negative few-shot sets from those annotations
  • Applies tracker-specific inference to new meetings
  • LLMs
  • AstraDB
  • Few-shot
  • Semantic Retrieval
06Visa

Customer 360 with RAG

LLM · RAG · Analytics

Ask the customer base a question in plain English.

A LangChain-powered dashboard that answers natural-language questions about user profiles and segments, removing the need to write SQL for business queries.

  • Retrieval-Augmented Generation over customer data with OpenAI GPT-4.1
  • Natural-language user profiling and segmentation
  • LangChain
  • OpenAI GPT-4.1
  • RAG
  • SQL
07Visa

Large-Scale Name Matching

NLP · Entity resolution

Finding one vendor among billions, typos included.

A scalable pipeline that searches billions of vendor records despite spelling errors and abbreviations.

  • Fine-tuned Sentence-BERT with triplet loss for name embeddings
  • Fuzzy matching applied only to the top-K retrieved candidates, improving query latency
  • Sentence-BERT
  • Triplet Loss
  • Embeddings
  • Fuzzy Matching
08PayU

Lead Prioritization

Machine learning · Fintech

Call the leads that will convert first.

CatBoost models that identify premium-quality leads for an online used-car reseller and an online gaming platform.

  • Achieved 3.7x lift for the used-car reseller and ~2.1x lift for the gaming platform
  • Trained and deployed end to end on AWS
  • CatBoost
  • Python
  • Pandas
  • AWS EC2
09PayU

SMS Income Estimation

NLP · NER · Lending

Estimating income from transaction messages.

Uses BERT named-entity recognition to extract amounts from SMS and classify each message as debit or credit.

  • Improved accuracy by 2.8% and coverage by 35% over traditional methods
  • BERT
  • NER
  • Python
  • HuggingFace
10PayU

Indian Address Standardization

NLP · Multilingual · Fine-tuning

Structure out of unstructured Indian addresses.

A multilingual model that parses free-text addresses into structured components and copes with spelling errors and partial addresses.

  • IndicBART fine-tuned with PEFT / LoRA
  • IndicBART
  • PEFT / LoRA
  • PyTorch
  • HuggingFace
11PayU

Propensity-to-Purchase

Machine learning · Predictive modelling

Who buys next month, across industries.

A reusable framework that predicts next-month purchase propensity for grocery, food, fashion and other merchants.

  • Mean model performance of 92.7% AUC and 65 KS
  • Python
  • SQL
  • Pandas
  • Gradient Boosting
12Jekson Vision

Tablet Inspection System

Computer vision · Industrial

Inspecting pharmaceutical tablets at line speed.

Computer-vision inspection of pharmaceutical tablets, optimised on GPU and extended with deep learning.

  • Increased inspection speed by 39% with GPU computer vision
  • Raised accuracy by 7% using deep-learning algorithms
  • Python
  • C++
  • OpenCV
  • TensorFlow
  • CUDA
13MINRO Lab, IIIT Bangalore

EEG Analysis with Deep Learning

Research · Signal processing · Vision

Reading brain signals as 1D, 2D and video.

MS research on deep-learning analysis of EEG, treating signals as sequences, images and video, including Parkinson’s disease detection and visual creativity analysis.

  • Four peer-reviewed publications across IEEE, ACM and Frontiers
  • Python
  • TensorFlow
  • Keras
  • OpenCV